Phi-4 Multimodal

Phi-4 Multimodal is a Phi-family model useful for small language model, edge, and low-resource workflow evaluation.

Microsoft · Phi

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesExact model card

Model checkpoints, context windows, provider support, local runtime compatibility, and license terms can change quickly. Verify the exact model card before production or commercial use.

Best for

Small multimodal and edge experiments

Who should use it

Small multimodal and edge experiments
Builders who want local or self-hosted testing options.

Common workflows

Small multimodal and edge experiments
small workflows
edge workflows
local workflows
efficient workflows

Deployment and hardware notes

~3.4 GB at Q4_K_M (5.6B parameters), so a single 8 GB card covers the whole text-plus-image-plus-audio pipeline that would otherwise need two models resident.

License and usage notes

MIT. Open weights where released. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

Open weights where released model option for Phi workflows.
Small multimodal and edge experiments
The reason to choose it is architectural, not benchmark-driven: one model to load, one to update, and no orchestration between a transcriber and a vision model.

Limitations

Folding vision and speech into 5.6B parameters costs depth in each: a dedicated Whisper checkpoint transcribes better, and larger vision-language models read documents better. Custom modeling code is required, and runtime support is thinner than for text-only Phi.
~3.4 GB at Q4_K_M (5.6B parameters), so a single 8 GB card covers the whole text-plus-image-plus-audio pipeline that would otherwise need two models resident.
Context window and limits: 131,072 tokens.
Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

The reason to choose it is architectural, not benchmark-driven: one model to load, one to update, and no orchestration between a transcriber and a vision model.

Local runtimes: Ollama where supported, LM Studio where supported, llama.cpp where supported, Transformers

Platforms: Windows, macOS, Linux

Vision spec

Memory~3.4 GB at Q4_K_M (5.6B parameters)Image inputDynamic cropping, up to 36 crops; also accepts audioContext131,072 tokens

One checkpoint, three input types: text, images and audio, with speech recognition and speech translation trained in rather than delegated to a separate model — which is what separates it from every other entry on this page.

Sources to verify

Exact model card

Related resources

Continue with model source notes, local tools, and implementation guides related to this model.

Phi family hub AI tools directory Model source notes Models directory Guides library

Hardware~3.4 GB at Q4_K_M (5.6B parameters)RuntimeOllama or LM Studio where supported, llama.cpp, Transformers, vLLMContext131,072 tokensLast updated2026

Exact model card →